Does Financial Market Structure Impact the Cost of Raising. Capital?

Transcription

1 Does Financial Market Structure Impact the Cost of Raising Capital? James Brugler, Carole Comerton-Forde and Terrence Hendershott * February 1, 2018 Abstract We examine the impact of secondary market structure and liquidity on the cost of raising capital. In the 1990s trading on Nasdaq transformed from a dealer-oriented over-the-counter market to a market where investors could directly interact with each other. The Order Handling Rules (OHR) reforms that accomplished this were phased in across Nasdaq stocks over time allowing for identification of their impact on firms cost of capital. We find that OHR significantly reduced the underpricing of seasoned equity offerings by one to two percentage points from a pre-ohr average of 3.6 percent. Using the staggered introduction of the OHR as an instrument shows improved secondary market liquidity drives the reduction in underpricing. * Brugler is at the Department of Finance, University of Melbourne. Comerton-Forde is at the School of Banking and Finance at the UNSW Business School. Hendershott is at the Haas School of Business, University of California. We thank Jeff Smith at Nasdaq for assistance in identifying and understanding the implementation of the Nasdaq reforms. We also thank Frank Hatheway, Ross Levine, Dan Li, Joel Hasbrouck and participants at the FINRA and Columbia University Market Structure Conference, the Australian National University Research Summer Camp and the NBER Conference on Competition and the Industrial Organization of Securities Markets for helpful comments. 1

2 1 Introduction Financial markets facilitate trading and price discovery by linking investors and firms. This impacts firms cost of raising capital and, consequently, their investment decisions. The market structure of trading affects secondary market liquidity (Madhavan, 2000). However, there is limited evidence on how the structure of the secondary markets impacts the cost of raising capital in the primary market through financing frictions. The cost of raising capital incorporates frictions in the issuing process, such as the explicit fees and implicit costs related to underpricing at issuance. 1 Our paper provides direct, causal evidence on the relationship among the underpricing of seasoned equity offerings (SEOs), market structure, and liquidity. Conducting such inference is complicated by a number of sources of endogeneity between underpricing and liquidity. 2 To estimate the effects of liquidity and market structure on SEO underpricing, we exploit a significant change in the market structure of stocks trading on the Nasdaq in 1997, the Order Handling Rules (OHR). These rules were designed to move Nasdaq from a dealer-oriented overthe-counter (OTC) like structure to a more centralized order-driven market structure. The OHR reforms were prompted by anti-competitive dealer behavior (Christie and Schultz, 1994). Two of the most important components of these rules were that dealers were required to display limit orders posted by members of public whenever these were at the best bid or offer and that dealers were required to publicly display their best quotes, rather than on segmented marketplaces that only certain types of investors could access. Barclay et al. (1999), McInish, Van Ness, and Van Ness (1998), Weston (2000), and Chung and Van Ness (2001) demonstrate that the OHR had the desired effect with quoted and effective spreads declining by about one third. While depth at the best bid and offer also fell, Conrad et al. 1 Corwin (2003), Butler, Grullon, and Weston (2005) and Ellul and Pagano (2006) examine underpricing and liquidity. Corwin (2003) estimates a positive but statistically weak association between bid-ask spreads and the underpricing of seasoned equity offerings. Butler, Grullon, and Weston (2005) show that stock liquidity is associated with lower fees charged by investment banks for seasoned equity offerings. Ellul and Pagano (2006) find that the expected level of liquidity and liquidity risk are determinants of IPO underpricing. 2 A key reason why endogeneity is present in regressions of underpricing on liquidity is the presence of an unobservable omitted variable, such as information asymmetry, which can directly affect both underpricing (Rock, 1986; Beatty and Ritter, 1986; Carter and Manaster, 1990) and liquidity (Copeland and Galai, 1983; Glosten and Milgrom, 1985; Kyle, 1985). Uncertainty and information asymmetry can also cause sample selection bias, either through their correlation with firms financing needs or due to firms strategically timing their issues to minimize costs. In addition, high underpricing may itself be a signal for liquidity providers that investors are unwilling to take large positions in a given security and therefore may be indicative of inventory risk, implying simultaneity that would bias simple least squares regressions. 2

3 (2003) show that the OHR decreased the cost of executing large institutional orders by 5 to 15 basis points from a pre-ohr average cost of 40 to 50 basis points. Crucially for our purposes, the OHR was specifically targeted at improving competition in market making on the Nasdaq. It plausibly did not have a direct effect on underpricing. Hence, in addition to examining the effect of the market structure change, we use the implementation of the rules as a source of exogenous variation in liquidity. There are a number of other features of the OHR that are attractive for our purposes. First, the rules were implemented in a staggered fashion, across 22 distinct dates covering a 10 month period in Second, while the cohort of stocks included in a given wave was determined by relative trading volume, there was a large degree of randomization within broad categories of stocks. Third, by the end of the implementation period, all stocks listed on the Nasdaq were covered by the rules implying that coverage went from zero to 100% in our sample period. Fourth, the OHR only affected Nasdaq stocks so we can use New York Stock Exchange (NYSE) stocks as controls. While the staggered introduction provides treatment and control samples for within Nasdasq study, the order of the stocks in the implementation of the OHR was not truly randomized. Stocks with higher relative trading volume more likely to enter the program earlier. As such, our empirical approach must be careful to distinguish between changes in liquidity that were due to the OHR and those that were simply due to different characteristics across stocks in different phases. Figures 1a and 1b demonstrate this point. Underpricing and bid-ask spreads were both lower for stocks completing SEOs after they were phased into the OHR. However, this may not reflect only a causal effect of the OHR on liquidity and underpricing, but also systematic differences in characteristics across OHR vs. non-ohr stocks. Figure 1 about here We use the OHR in several ways. First, we treat the OHR as a quasi-random experiment and estimate its direct effect on SEO underpricing in a pooled difference-in-differences framework. These regressions treat the OHR status as a dummy variable and estimate its effect both with and without controls, stock-cohort fixed effects based on the date of inclusion in the OHR and time fixed effects. By pooling our estimates across dates, we are able to get an estimate of the treatment 3

4 effect of the OHR without being so reliant on the parallel trends assumption as with only a single treatment date (Bertrand and Mullainathan, 2003; Gormley and Matsa, 2011). In our second approach, we use the OHR as an instrumental variable for liquidity in a regression with SEO underpricing as the dependent variable. These regressions complement the difference-in-differences regressions by allowing us to directly test whether any effect of OHR on SEO underpricing was due to the influence that the new trading rules had on liquidity, rather than through some other channel. These regressions also allow us to directly estimate the marginal response of capital costs in the form of SEO underpricing to changes in market liquidity. We find that the Order Handling Rules had a statistically and economically significant effect of reducing SEO underpricing. In a difference-in-differences specification that includes cohort and time fixed effects as well as stock and issue controls, SEOs of companies with stock trading under the OHR were less underpriced by 2.18%, as compared to a 3.6% pre-ohr average SEO underpricing. Our instrumental variable regressions confirm this result and show that improved secondary market liquidity is the channel by which the OHR reduces underpricing. Using the variation in liquidity that is driven by the OHR, we find that lower stock liquidity leads to higher SEO underpricing. This effect is both statistically and economically significant. Further, the magnitude of the effect estimated in these regressions is very similar to that estimated in the difference-indifferences approach, when appropriately scaled. We interpret these results together as supportive of the notion that the effect of the OHR on underpricing is due primarily to changes in liquidity, and not due to some other factor that we have not controlled for. In our third approach, to eliminate any concerns about non-random assignment in the rollout schedule for OHR stocks we use NYSE stocks as controls to estimate the impact of the OHR on SEO underpricing on Nasdaq. Because Nasdaq stocks are smaller and more volatile than NYSE stocks we match SEOs across exchanges based on their issuers characteristics. Figures 2 and 3 plot the average underpricing and pre-issue bid-ask spread for Nasdaq and NYSE SEOs from June 1996 to June Panel (a) of these figures plots the means for all SEOs on both exchanges and Panel (b) contains the mean for all Nasdaq SEOs and the mean for the matched sample of NYSE SEOs where matching is conducted as per Section 4.3. Figures 2 and 3 about here 4

5 Figures 2 and 3 show a reduction in the gap between Nasdaq and NYSE underpricing and preissue bid-ask spreads from before the implementation of the OHR to after the implementation of the OHR. The average Nasdaq SEO underpricing decline from 2.8% more than the average NYSE SEO to approximately 1.5% more than the average matched NYSE SEO. For average SEO bidask spreads, the Nasdaq-NYSE difference is roughly 1.5% and 0.5% for raw and matched NYSE samples respectively. Post-OHR this gap is eliminated for the raw sample and reversed for the match sample. The remainder of this paper proceeds as follows. In Section 2, we describe the Order Handling Rules in detail and discuss previous work relating it to liquidity. In Section 3 we describe data sources and provide summary statistics. Section 4 discusses our empirical approach for identifying the effect of liquidity on SEO underpricing using the OHR as a source of exogenous variation in liquidity. Section 5 describes our results. Section 6 summarizes our findings. 2 Institutional Background The OHR comprise a number of changes. Weston (2000) emphasizes that the OHR reduces the OTC-like nature of Nasdaq and increases competition in liquidity supply in two main ways. 3 First, the Limit Order Display Rule requires market makers to display investor limit orders if they are priced better than the market maker s quote. This rule enables investors to compete against dealers for order flow, and enables investors to access limit orders that were not previously displayed to the market. Second, the Quote Rule requires market makers to publicly display their best quotes. Market makers had been previously able to post different quotes on Nasdaq and on Electronic Communications Networks (ECNs), which are not universally accessible. The OHR were implemented using a staggered phase-in with 22 waves. The first wave of stocks began on January 20, 1997 and the last wave began October 13, The first 13 waves included the Top 1000 Nasdaq stocks by median dollar volume, with each wave including the 10 largest volume stocks and a random draw of eight stocks from the top five deciles. Wave 14, which began 4 August is the first wave from which stocks were drawn from the entire Nasdaq universe. The initial waves comprised only 50 stocks, but the majority of stocks are phased-in in large groups 3 Other changes in the OHR include a reduction in the Minimum Quote Size from 1,000 shares to 100 shares and the relaxation of the Excess Spread Rule. 5

6 of approximately 850 stocks during September and the first half of October. A summary of the number of stocks phased-in in each wave is provided in Figure 4. Further details about the roll-out are provided in Smith (1998). Figure 4 about here Another major change to the Nasdaq market structure occurred during the OHR roll-out period. On May 27, 1997 the SEC approved a reduction in tick size from $1/8th to $1/16th. This change was implemented June 2, Hatheway and So (2006) describe how on March 4, 1997 the SEC adopted Regulation M ( Reg M ) which eased restrictions on passive market making for underwriters during the five days leading up to the offering. Because underwriting investment banks were often market makers in the stock, pre-reg M limits on their market making could impact prices and liquidity prior to the SEO. This could impact SEO underpricing. Therefore, we also examine underpricing on Nasdaq by OHR status for the post-tick-size change period. The OHR had an immediate and dramatic impact on Nasdaq market quality. In a study of the first two phase-in samples Barclay et al. (1999) show that the quoted and effective spreads declined by approximately one third. They report that spreads decline for all stocks, but decline by a larger magnitude in less active stocks, and for stocks with large pre-ohr spreads. The results for depth are ambiguous. When ECN quotes are included in depth calculations, inside depth is unaffected. However, this fails to consider that ECN depth is available before the OHR, but not captured in the data. When ECN depth is excluded, the first phase-in sample stocks exhibit a decline in depth, but the second phase-in sample exhibits an increase in depth. This likely reflects the fact that the Minimum Quote Size is reduced for the first phase-in sample, but not the second. Barclay et al. (1999) and McInish et al. (1998) also show that average trade size declines, but the number of executions increases. Smith (1998) examines the complete implementation of the OHR and the reduction in tick size from eighths to sixteenths. He argues that these two changes affected quoted spreads differently. The spread declines are larger for more active stocks, but lower for higher-priced, larger and more volatile stocks. For the most active stocks, the OHR in isolation did not impact spreads as these 4 This change applied to stocks priced above $10. For stocks priced less than $10, quotes could be expressed in increments of $1/32, before and after this change. 6

7 stocks traded at the lower bound of $1/8th. Smith also shows that the inside spread is more likely to be set by orders placed in ECNs in active, high-priced stocks. Depth results for the full sample are also mixed, with high-priced stocks exhibiting greater declines/smaller increase in depth compared to lower-price stocks. 3 Data and Summary Statistics SEO and issue characteristics are obtained from the Securities Data Company (SDC) New Issues database. For our main sample of SEOs on the Nasdaq, we use a sample period from January 1997 to October 1997 inclusive, which covers the entire roll-out of the OHR. For our analysis that compares Nasdaq and NYSE SEOs, we use data covering the period June to December 1996 and January to June 1998 (i.e. six months of data in the year before and after the implementation of the OHR). Similar to Lee and Masulis (2009) and Karpoff et al. (2013), we include SEOs of common shares by public US companies with an offer price of at least $5, sold on a firm commitment basis and exclude rights issues and depository receipts. Sales by real estate investment trusts are excluded as are issues with a filing date of more than 12 months before the beginning of our sample. 5 These filters yield a total of 213 Nasdaq SEOs that meet our criteria between January and October For each SEO we observe the 9-digit CUSIP, the stock ticker symbol, the issue date as determined by SDC, the offer size (in $ millions) and the offer price. For each stock in our sample, we obtain CRSP daily data containing the closing price, best bid and ask, volume traded and shares outstanding. From these data we construct the average closing bid-ask spread as a percentage of mid-quote price in the month prior to the issue date for each SEO. We refer to this as the bid-ask spread and we use this as one of our two liquidity variables. We also use the CRSP data to construct control variables including the log of market capitalization, the log of stock price, the standard deviation of one month daily returns (volatility) and monthly volume traded (in $ millions). These controls are similar to those used by Corwin (2003). We use the method of Safieddine and Wilhelm (1996) to adjust the issue date for SEOs that occur after the close of trading. Safieddine and Wilhelm (1996) use spikes in trading volume to identify the actual SEO issue date. If the day following the stated issue date has at least twice the 5 Only two issues meeting the other filters had filing dates prior to Jan 1,

8 trading volume of the stated issue date, then the issue date is adjusted to be the next trading day. Corwin (2003) and Karpoff et al. (2013) both use this method to identify the correct issue date. For our Nasdaq regressions covering the actual roll-out of the OHR, market capitalization and price are determined using the most recently recorded entry observed no later than January 2nd, Volatility and volume are recorded over the month of December, We define our control variables at this point in time (prior to the initiation of the OHR) to ensure that our inference is not biased by any indirect effects that the OHR may have on the control variables, for example via volume traded or price. Of our initial sample of 213 SEOs that meet our criteria, 12 do not have CRSP data available as at January 1, These companies issue an SEO at some point in our sample, but are yet to IPO by the date at which we define our control variables. We also run each of our regressions using covariates defined as the day or month immediately preceding the issue date and obtain very similar results. For the analysis comparing Nasdaq and NYSE SEOs before and after the OHR roll-out, we record the same control variables calculated using data from the most recent month prior to the SEO issue date. We construct our underpricing dependent variable using the issue price in the SDC data and the closing price in CRSP on the day prior to the issue date. Underpricing is defined as per the main regressions in Corwin (2003) and is the negative of the log return from the previous closing transaction price to the offer price in percentage terms. We also construct issue relative size, defined as the value of the issue divided by the market capitalization. In addition to bid-ask spreads, we estimate stock-level liquidity using the Corwin and Schultz (2012) spread estimate constructed from daily high and low prices (which we refer to as the high-low spread). The logic behind the high-low spread estimate is that the ratio of the daily high to daily low price is determined by the stock s variance and its bid-ask spread. The variance component is proportional to the return interval while the spread component is not. From these simple facts, Corwin and Schultz (2012) construct an estimate of spreads from high-low price ratios over oneand two-day intervals. Monthly estimates of the high-low spread are available on Shane Corwin s personal web-page (Corwin, 2017) and we use these data as our second liquidity variable. 6 All variables stock and issue characteristics are Winsorized at the 1% level, except for the log of stock 6 Liquidity estimates based on the monthly Amihud ratio and the bid-ask spread in dollars generate qualitatively similar results to those using the CRSP bid-ask spread and the high-low spreads. These results are omitted from this version of the paper but are available upon request. 8

9 price. The implementation schedule for the OHR was obtained from two sources: Nasdaq equity trader alerts during the 1997, published via Nasdaq (2017) and a proprietary list of inclusion dates provided to us by Nasdaq. 7 The trader alerts are in PDF format and cover the period January 1, 1997 onwards. These alerts were issued to market participants usually one to two weeks in advance of each phase of the implementation schedule. They contain the ticker symbol for each stock included in each of the 22 phases of the implementation from Wave 2 (February 10, 1997) onwards. The list of inclusion dates provided by Nasdaq also contains stock tickers and implementation dates, and also covers the first 50 stocks included in the pilot program implemented on January 20, These data provide us with the date that each stock was included in the OHR and we match to SEOs by ticker symbol. From these dates we construct a dummy variable indicating whether a company s stock is trading under the OHR at the date of the SEO (value 1) or not trading under the OHR at the date of the SEO (value 0). 8 We are able to match all but five of our SEOs by ticker symbol into the Nasdaq data which leaves a total of 208 SEOs for which we observe issuing characteristics and OHR phase-in date and 196 ( ) for which we observe issuing characteristics, OHR phase-in date and CRSP control variables. The number of stocks included in each wave of the phase-in schedule is plotted in Figure 4. This figure shows that the majority of stocks were not phased in until August and September. Figure 5 plots the number of SEOs by phase-in status (i.e. delineated by whether the stock was trading under the OHR or not) by month. Consistent with Figure 4 we observe that until the end of July 1997 most SEOs are done by companies with stocks not trading under the OHR. After this time we observe the number of SEOs done by OHR companies rise and non-ohr companies fall, until October 1997, at which time all stocks were included in the program. Figure 5 about here Table 1 contains summary statistics of our data. The mean SEO underpricing in our sample is 7 We gratefully thank Jeffrey Smith for his assistance with providing this list. 8 While there is a very high degree of consistency across the two datasets, we use the Nasdaq trader alerts where possible as these are considered to be the most official record available according to Nasdaq economists. 9

10 2.98% with a standard deviation of 3.21%. The median underpricing is 2.03%, the average SEO represents 27.1% of the current market capitalization of the firm, the average bid-ask spread is 2.30% and average one month standard deviation of returns is 3.00%. The equivalent Nasdaq averages from Corwin (2003) are 2.72% for close to offer underpricing, 26.84% for relative size, 2.95% for bid-ask spread and 3.41% for one month standard deviation of returns. The data used in Corwin (2003) covers 1980 to 1998 for the issuing characteristics and 1993 to 1998 for liquidity. Table 1 about here Table 2 contains correlation coefficients for our data. Underpricing is negatively correlated with the value of the issue and market capitalization but positively correlated with relative size, indicating that SEOs by larger companies tend to have less underpricing but that participants in SEOs receive more compensation when companies raise relatively more capital. Volatility and illiquidity (bid-ask and high-low spreads) are both positively correlated with underpricing while the OHR dummy variable is negatively correlated with underpricing. The negative correlation between the OHR dummy and underpricing at least partially reflects the fact that larger stocks were phased in earlier in the program, as well as any potential causal relation running from liquidity to underpricing. Table 2 about here Table 3 contains summary statistics over the period June 1996 to June 1998 split by exchange. The second column of Table 3 contains the mean for Nasdaq issues, the third column contains the means for all NYSE issues and the fourth column contains the t-statistic for a test of equal means between Nasdaq and NYSE issues. The fifth column contains the means for a nearestneighbour matched sample of NYSE issues where matching is conducted on issue date, log of market capitalization, log of stock price, volume traded, volatility and issue size using the inverted variance weighting matrix (full details in Section 4.3). Table 3 indicates that SEOs on the Nasdaq tend to be smaller but represent a larger fraction of existing equity capital and are more significantly underpriced. The matching process reduces the gap between average Nasdaq and NYSE SEO 10

11 characteristics, but is unable to entirely eliminate the differences. 9 There were also substantially more SEOs taking place on the Nasdaq over our sample period than on the NYSE, as shown in Figure 6. Table 3 about here 4 Identification using the Order Handling Rules Our approach to identifying the effect of liquidity on SEO underpricing exploits the change in market structure following the introduction of the OHR as a shock to liquidity across Nasdaq securities. The OHR reforms were designed to offer investors more competitive quotes via the mandatory display by dealers of superior customer limit orders and the dissemination of superior prices posted on proprietary trading venues such as ECNs. Subsequent studies by Barclay et al. (1999), McInish et al. (1998) and Chung and Van Ness (2001) demonstrate that the OHR led to a statistically and economically significant reduction in spreads (quoted and effective). For our purposes, we use the OHR as a quasi-natural experiment that drives variation in stock liquidity but that does not directly affect underpricing and is unrelated to potential omitted variables such as information asymmetry. While the staggered introduction of the OHR was not entirely random, (the first 13 waves were only drawn from the 1000 most actively traded issues), there was a large degree of randomization within each wave. Indeed from August 4 th onwards, stocks were drawn randomly from the entire universe of Nasdaq issues (Smith, 1998). Furthermore, selection into each wave was determined by relative trading activity (volume). Given the assignment to waves on observables (trading volume) and the randomization within each wave combined with the market-wide implementation of the new market structure, we argue that the OHR is appropriate for our purposes. 4.1 Regression Specifications: Differences-in-Differences We use the OHR dummy variable in three related econometric models. First, we use OHR status as a treatment variable and estimate the effect of OHR status on SEO underpricing in a difference- 9 We also use the Mahalanobis weighting matrix and the Euclidean weighting matrix but these generate worse matches than those using the inverted variance weights. 11

12 in-differences framework. The most general version of these regressions takes the form: y it = γ c + µ t + βohr it + ρ x it + ε it (1) where y it is the underpricing of the i th SEO during time period t, OHR it is the OHR status of the issue (1 if trading under the OHR at time of issue and 0 otherwise), x it is the vector of stock-specific control variables defined in the period before the the start of OHR implementation, γ c is a fixed effect defined by membership of each of the phase-in waves (i.e. γ j takes the value 1 if stock i was included in the j th wave of stocks) and µ t is a time fixed effect where time is defined either as calendar month or by the series of dates at which new stocks were introduced to the OHR (i.e. ρ j takes the value 1 if the issue occurs in the j th month or between the OHR inclusion dates of the j th and j + 1 th waves, depending on how the time fixed effects are being defined). With wave-cohort fixed effects and time fixed effects defined by the dates of each wave s introduction to the OHR, Equation (1) is analogous to a treatment effect around a single treatment date, but where assignment to treatment or control occurs across multiple groups and periods. A similar approach is used in both Bertrand and Mullainathan (2003) and Gormley and Matsa (2011) and is also applied in the context of corporate bond issuing costs and transparency by Brugler, Comerton-Forde, and Martin (2016). As discussed in Brugler et al. (2016), the parameter β is our pooled analogue of the coefficient on the interacted term between the treatment dummy and the post-treatment period dummy in a difference-in-difference model using a single treatment period. It captures the average treatment effect across the multiple events. Pooling the 22 treatment dates into a single regression allows us to control for cohort-specific effects and means we are not as reliant on the parallel trends assumption as we would be when analyzing the difference-in-difference around a single event. We estimate Equation (1) under four specifications: excluding controls and fixed effects (i.e. regressing underpricing only on OHR status), including the controls, including controls and monthly fixed effects, and including controls, wave-cohort fixed effects and time fixed effects based on wave dates. As per Secton 3, the control variables in the relevant specifications are log of market capitalization, relative size of SEO, volatility of mid-quote returns, log of stock price and log of volume traded, defined in the period prior to the initial roll-out of the OHR where applicable. 12

13 Of course, implementation of the OHR is not truly random. If it were, arguably the most rigorous way to estimate Equation (1) would be to exclude all control variables as inclusion of the wave-cohort fixed effects can theoretically remove any time-invariant stock characteristics that may affect SEO underpricing and differ systematically across cohorts. The fact that OHR status is driven in part by relative trading volume motivates us to incorporate the controls, however estimating the model without controls and only wave-cohort dummies does not affect our conclusions. 10 We also estimate Equation (1) for three sub-samples of our data. The first two sub-samples are based on market capitalization: we estimate Equation (1) for SEOs by companies with market capitalizations below the sample median and for SEOs by companies with market capitalizations above the sample median. These regressions are designed to allow for heterogeneous treatment effects for OHR status between smaller and larger stocks. The third sub-sample includes all SEOs that take place after June 2nd, These regressions are provided to address potential concerns about the implementation of two additional trading rule changes that affect all Nasdaq stocks in 1997: the change in tick size from 1/8th to 1/16th on June 2nd, 1997 and Regulation M that impacted the actions that deal participants could undertake around new securities offerings and was implemented on March 4th, Although date fixed effects should theoretically account for the market-wide impact of these changes, we include the regressions on a sub-sample where trading rules were unchanged as an additional robustness check. 4.2 Regression Specifications: Instrumental Variables The second way in which we exploit the OHR is as an instrumental variable (IV) for liquidity in two-stage least squares (2SLS) underpricing regressions. The target regression model we wish to estimate is: y it = µ t + βliq it + ρ x it + ε it (2) where Liq it is the liquidity (bid-ask or high-low spread) of stock i undergoing an SEO at time t where liquidity is the averaged across the month preceding the issue date. Other variables are defined as per Equation (1). Due to omitted variable bias and selection on unobservables, Liq it is potentially correlated with the error term ε it. Our solution to this problem is to instrument for 10 These results are available upon request. 13

14 Liq it using the OHR status of the stock being issued. The relevance requirement for our instrument is that the implementation of the OHR is associated with an economically and statistically significant improvement in liquidity, and specifically a reduction in spreads. Consistent with the evidence presented in Barclay et al. (1999), McInish et al. (1998) and Chung and Van Ness (2001) our first stage regression results show that this result also holds in our sample. The exogeneity condition requires that, conditional on relevant control variables, our instrument only affects underpricing through the liquidity channel, and (1) does not directly drive underpricing itself or (2) affect underpricing through any other channel that is not controlled for. Our structural estimates of the parameters in (2) are only just-identified, so we cannot provide evidence via overidentifying restrictions, such as with a Sargan or Hansen J-test. Instead we must rely on the pseudo-random nature of the OHR implementation schedule combined with the fully observable nature of assignment to waves (based on trading activity) to justify the validity of our instrument. For all models and specifications, we calculate White heteroskedasticity-robust standard errors and report tests based on these standard errors. We have also estimated all models with standard errors clustered at the time level. Cameron and Miller (2015) note that parameter covariance matrices can be downward biased when there are few clusters and that this problem can be particularly problematic when the number of observations by clusters varies. Given the highly unbalanced nature of the clusters in our sample and the relatively few clusters (either 10 or 23 depending on how the time fixed effects are defined), we rely on our simple White standard errors. However, our conclusions are not sensitive to clustering by time. 4.3 Regression Specifications: Comparisons with SEOs on the New York Stock Exchange Under the two specifications outlined in Sections 4.1 and 4.2, we use differences in OHR status across Nasdaq stocks to identify the effect of market structure on capital costs. Although the OHR provides us with a source of variation in stock liquidity that does not directly affect underpricing, the nature of the roll-out of the program implies that OHR status is not truly random. As an alternative to including OHR status directly as a regressor or instrument in an econometric model, 14

15 we instead analyse how underpricing and liquidity changed before and after the implementation of the OHR for SEOs taking place on the Nasdaq exchange relative to a group of SEOs for which no major change in market structure takes place, namely those that occur on the NYSE. We compare Nasdaq and NYSE SEOs in a number of ways. First, we do a simple comparison of means for underpricing and liquidity (bid-ask spreads and high-low spreads) between Nasdaq and NYSE SEOs before the beginning of the implementation of the OHR. We then compare the means for underpricing and liquidity between Nasdaq and NYSE SEOs after the implementation of the OHR is complete and calculate the difference-in-differences in means. This difference-in-difference provides a simple estimate of the degree to which SEO underpricing improved or deteriorated on the venue where the OHR was implemented relative to a control venue with no change in market structure. It also is identical to the treatment effect coefficient for a regression of underpricing or liquidity on exchange dummies, time dummies for whether the SEO occurred before or after the OHR implementation and an interaction term between exchange and time dummies: y i = δ 0 + δ 1 t i + δ 2 Nas i + τ t i Nas i + ε i (3) where y i is the underpricing or liquidity of the i th SEO, t i is a dummy variable indicating whether the i th SEO occurred after the OHR implementation, Nas i is a dummy variable for SEOs on the Nasdaq. An alternative way to calculate this difference-in-differences is to compare the change in mean underpricing and liquidity for Nasdaq SEOs before and after the OHR implementation and compare this to the change in NYSE SEOs over the same period. The difference-in-differences from doing so is the same as calculating the differences-in-differences between Nasdaq and NYSE SEOs within a given period, however this alternative representation allows us to isolate whether SEOs on the Nasdaq or the NYSE are responsible for the results. Interpreting the simple comparison of means between Nasdaq and NYSE SEOs is complicated by systematic differences in characteristics of stocks listed on the two exchanges. Table 3 demonstrates that companies undertaking SEOs on the NYSE in our sample period are larger and issue larger amounts of stock. They also have higher prices and volume, better liquidity and lower volatility than SEOs on the Nasdaq. Since we are comparing changes in means, these differences in average 15

16 characteristics do not necessarily invalidate our approach, however if trends in SEO underpricing differ across stocks with different characteristics, then inference from this approach will be biased. To deal with this, we again estimate a treatment effect model where treatment status is defined by exchange status, with Nasdaq stocks as the treated group but we now include the control variables capturing differences in SEO and stock characteristics. The estimated model is given by: y i = δ 0 + δ 1 t i + δ 2 Nas i + τ t i Nas i + β X i + ε i (4) where y i, t i and Nas i are defined as in Equation (3) and X i is the set of control variables. These are issue date, SEO relative size, log of market capitalization, log of stock price, volume traded in month prior to SEO and volatility of mid-quote returns in month prior to the SEO. Next, we compare the difference in average underpricing or liquidity between Nasdaq SEOs and a matched sample of NYSE SEOs both prior to the implementation of the OHR and following the implementation of the OHR. Each Nasdaq SEO is matched to its nearest neighbour NYSE SEO from the same period (pre- or post- OHR) where matching is conducted on the same SEO and stock characteristics as used in the estimation of Equation (4). From these differences across Nasdaq and matched NYSE SEOs within periods, we construct a matched difference-in-differences estimator in the spirit of of Heckman, Ichimura, and Todd (1997), Heckman, Ichimura, Smith, and Todd (1998) and Todd (2010). As per Todd (2010), this estimator for repeated cross-sections takes the form ˆα 1 = 1 { NP Nas Y Nas ost P ost,i j I NYSE Post W (i, j)y NY SE P ost,j } 1 { NP Nas Y Nas re P re,i j I NYSE Pre W (i, j)y NY SE P re,j } (5) where Y e t,i is the outcome variable (underpricing or liquidity) for the ith SEO, occurring in period t {P re, P ost} on exchange e {Nasdaq, NY SE}, N Nas time the Nasdaq in period t, I NYSE t is the number of SEOs occurring on is the set of indices of SEOs that occur on the NYSE in period t {P re, P ost} and W (i, j) is a weighting function that takes the value 1 if the j th NYSE SEO is the nearest neighbour match for the i th Nasdaq SEO and zero otherwise. In the treatment effect literature, estimators of the form (5) are used to estimate a treatment effect when systematic differences between participant and non-participant outcomes persist, even after conditioning, which may be due to selection on unobservables, for example (i.e. violations of 16

17 the strong ignorability condition of Rosenbaum and Rubin (1983)). In our approach, taking the difference-in-differences of the two matched estimators (i.e. before and after the OHR implementation) allows us to infer whether or not the effect of listing location on underpricing and liquidity changed significantly after the introduction of the OHR on the Nasdaq exchange, while accounting for differences in average characteristics which may change over time or imply different trends in outcome variables. We use the inverted variance weighting matrix to calculate each nearest neighbour as this generated a better fit than either the Mahalanobis or Euclidean weights. We obtain very similar results using these different weighting matrices but as the match quality is worse, we rely on the inverted variance weights. We lastly complement the estimation of (5) by estimating an additional matched difference-indifferences estimator where we first match Nasdaq (NYSE) SEOs that occur after the implementation of the OHR with Nasdaq (NYSE) SEOs that occur prior to the implementation of the OHR. We then calculate the difference in these two estimators. This estimator takes the form ˆα 2 = 1 { NP Nas Y Nas ost P ost,i j I Nas Post } W (i, j)yp Nas re,j N 1 NY SE P ost { NY SE YP ost,i j I NYSE Post W (i, j)y NY SE P re,j } (6) where all variables are defined as in Equation (5). Equation (6) is analogous to our comparison of the change in raw means from before and after the implementation of the OHR for Nasdaq SEOs and NYSE SEOs respectively, while accounting for potential changes in the average characteristics of SEOs across time. Standard errors for the difference in raw means are calculated without assuming identical variances between sub-samples. We use a non-parametric bootstrap for calculating standard errors in matched differences and White heteroskedasticity-robust standard errors for treatment effect regressions (Equations (3) and (4)). We include all SEOs that take place between June 1996 and Dec 1996 as our pre-ohr sample and those that take place between January 1998 and June 1998 as our post-ohr sample. 17

18 5 SEO Results 5.1 Difference-in-Differences Regressions We begin our analysis with a discussion of our differences-in-differences estimates that treat the Order Handling Rules as a quasi-random treatment effect and estimate its direct effect on underpricing. The parameter estimates and associated t-statistics for our pooled difference-in-differences estimates of Equation (1) are contained in Table 4. Model A contains estimates from a regression of SEO underpricing onto the OHR dummy and a constant term, without controls or fixed effects. Model B contains analogous estimates but with the inclusion of the control variables described in Section 4.1. Model C adds 10 calendar month time fixed effects to Model B. Model D includes 23 time fixed effects based on the roll-out dates of the OHR program and also cohort fixed effects for stocks in each wave of the OHR implementation schedule as well as the controls. Since this specification most closely adheres to a standard difference-in-difference framework, it is our preferred specification. Panel A of Table 4 contains results for all SEOs in our sample. Panel B contains results for stocks with above and below median market capitalizations respectively. Panel C contains results for regressions using only SEOs occuring after June 2nd, Table 4 about here For each of the models in Table 4, the OHR parameter is negative and significant at the 5% level or better, indicating that SEOs for companies with stock trading under the OHR were underpriced significantly less than SEOs of companies with stock yet to be phased into the program. In terms of economic significance, underpricing is predicted to be between 1.4% to 2.2% lower for stocks trading under the OHR compared with not for the full sample of SEOs, which represents between 47% to 68% of the sample standard deviation of underpricing. The magnitude of the parameter is largest for our most general specification in Model D. Firm size and price are negatively related to SEO underpricing and are significant at the 5% and 10% level depending on the specification. Offer size, volatility and volume are all positively related to SEO underpricing, but none are significant at the 10% level in any specification. The key take-away from Panel A of Table 4 is that, using a model with granular cohort and 18

19 time fixed effects and control variables that are known to be related to OHR status, we estimate that the Order Handling Rules led to a statistically and economically significant improvement in SEO underpricing and therefore a reduction in one source of direct capital costs for firms. Under this model, we can have a relatively high degree of confidence that conditional unconfoundedness holds. Furthermore, since the Order Handling Rules are applied to none of the SEOs at the start of our sample but apply to all SEOs by the end of it, with a significant period of overlap in the second half of 1997, there is little concern about weak overlap between the treated and control sub-populations in our data. Panel B of Table 4 indicates that that the magnitude of the effect of the OHR was larger for smaller stocks. For these stocks, the OHR led to a reduction in underpricing of between 1.6% and 3.5%. For larger stocks, the estimated effect of the OHR is between 0.9% and 1.4% depending on the specification. However, SEOs by companies with smaller market capitalizations tend to be more heavily underpriced than SEOs by larger companies. Comparing the size of the OHR parameters across small and large stocks to the mean underpricing by sub-sample shows that the standardized size of the effect is comparable across small and large stocks. For small stocks, the size of the effect is between 40% and and 80% of the sub-sample mean of underpricing (4.36%). For large stocks, the OHR coefficient represents betwen 55% and 80% of the sub-sample mean (1.7%). Nevertheless, it is plausible that it is the absolute change in the cost of issuing equity capital that matters for companies, not the relative change. If this is the case, then the results in Table 4 suggest that the OHR was relatively more beneficial for smaller companies than larger companies. The regressions in Panels A and B of Table 4 are estimated using data that spans the entire roll-out of the OHR. During this period, two additional rule changes affected the trading of Nasdaq stocks: the change in quotation tick size from one eighth to one sixteenth on June 2, 1997 and Regulation M that was effectively implemented on March 4, Panel C of Table 4 show that the OHR treatment effect is larger in the shorter sample, demonstrating that the possible confounding effects earlier in 1997 are not responsible for the OHR treatment effect. 19

20 5.2 Instrumental Variable Regressions As per Barclay et al. (1999), the OHR were specifically designed to offer investors more competitive quotes and the rules were effective in achieving these ends. Clearly, one obvious channel by which the OHR would drive changes in SEO underpricing is directly through liquidity. While it is not obvious to us what other channels may be directly affected by the OHR, instrumental variable regressions can help provide evidence that 1) the OHR did affect liquidity for the stocks in our sample and 2) the change in liquidity caused by the OHR was itself a key driver of the improvement in SEO underpricing. Parameter estimates and associated t-statistics for our IV regressions where liquidity is treated as an endogenous regressor and the OHR dummy as the IV are contained in Tables 5 and 6. In Table 5, the liquidity variable is the bid-ask spread while in Table 6, the liquidity variable is the Corwin and Schultz (2012) high-low spread estimator. In both tables, Model A includes only a constant term and the endogenous liquidity regressor that is instrumented for using the OHR dummy. Model B adds control variables while Model C adds controls and calendar month time fixed effects. Underneath the main regression estimates, we also report the first stage OHR dummy coefficient and t-statistic, the Kleibergen-Paap LM test for full rank of the first stage Z X matrix (underidentification), the Cragg-Donald Wald - F statistic for weak identification and the F -statistic for the relevance of the instrument in the first stage regression. Note that since there is a single instrument, the first stage F -statistic is simply the square of the t-statistic on the first stage OHR coefficient. Table 5 about here In all three specifications in Table 5, the second stage coefficient on the bid-ask spread is positive and both economically and statistically significant at the 5% level. The parameter is between approximately 50% and 300% larger than the equivalent OLS regressions, which possibly suggests that the OLS liquidity parameter estimates are biased towards zero. 11 This may explain to some degree the findings of Corwin (2003) that bid-ask spreads are only very weakly related to SEO underpricing. Our second stage estimates suggest that an exogenous one standard deviation 11 OLS regression results are available from the authors on request. 20

21 increase in bid-ask spreads would reduce SEO underpricing by between approximately 1.60% to 1.90%, or between about 50% to 60% of the sample standard deviation in underpricing. In all three specifications, we reject the null of underidentification and weak identification, and first stage F -statistics all exceed 10. An alternative way to assess the magnitude of our second stage regressions is to calculate the difference in expected underpricing for a stock with OHR status equal to one and one with OHR status equal to zero. To do this, we simply multiply the first stage coefficient with the second stage coefficient. Doing this yields expected changes in underpricing of between -1.40% to -1.55%. The magnitude of these effects are very similar to that of the OHR dummy variable in the equivalent difference-in-differences specifications in Table 4 and discussed in Section 5.1 (i.e. Models A, B and C in Table 4). One interpretation of the high degree of similarity between our IV results and our difference-in-differences results is that the causal effect of the OHR on SEO underpricing can be almost fully explained by the degree to which the OHR improved liquidity. We interpret the consistency in the magnitudes of effects across Tables 4 and 5 as supportive of the notion that OHR affected underpricing primarily through the liquidity channel. Table 6 about here Turning to our IV regressions that use the high-low spread estimator of Corwin and Schultz (2012), we find very similar results to those using the bid-ask spread. The second stage coefficients on our illiquidity variable are all economically and statistically significant at the 5% level or better. An exogenous one standard deviation increase high-low spreads would lead to a reduction in SEO underpricing of between 2.15% and 2.25%, or approximately two thirds of one standard deviation of underpricing in our sample. Calculating the magnitude of the effect for a stock trading in the OHR compared with one that is not included in the OHR, we again predict an effect that is very similar in size to those found in Models A, B and C of Table 4. We note that although the first stage diagnostic tests reject the nulls of underidentification and weak identification for Models A and B, for Model C we have some evidence of weak identification with a first stage F -statistic below 10 and the Cragg-Donald Wald test statistic below the relevant 10% Stock-Yogo critical values. In both Tables 5 and 6, the other control variables have very similar interpretations as the 21